What Knowledge Is

Entropy reduction across branching timelines

You glance at the hallway clock, read 4:17, and go on with your day believing it is 4:17. It is, in fact, 4:17. Your belief is true. It is also justified — consulting a clock that has served you faithfully for years is exactly what a responsible epistemic agent does. But the clock stopped last night, at 4:17, and you happened to walk past during the one minute in twelve hours when a dead clock tells the truth.

Do you know what time it is? On the classical definition of knowledge — justified true belief, a formula that goes back to Plato — you do. All three boxes are checked. And yet the verdict is obviously wrong: a true belief acquired by dumb luck is not knowledge, whatever boxes it checks. This is the Gettier problem, and since 1963 it has spawned an entire industry of patches — fourth conditions, “no false lemmas,” defeasibility clauses — each of which falls to the next round of counterexamples. When a definition needs a new epicycle every decade, the definition is wrong. I propose to replace it.

The Definition

Knowledge is pattern-encoded information in an agent’s predictive structure that reliably and quantifiably reduces entropy about specified future events or states across branching timelines.

Every word is doing work, so let me unpack them.

Pattern. Knowledge lives in a reproducible structure — a neural configuration, a logical rule, an algorithm, a cultural convention. The form does not matter; the stability does. A pattern too fragile or too idiosyncratic to be exercised twice cannot support prediction, and prediction is the job.

Predictive structure. The pattern must be installed in an agent — wired into the machinery that generates expectations and guides action. Knowledge is not a property of propositions floating in Plato’s heaven; it is a property of agents in the world, and it earns its title by what it lets those agents anticipate and do. The predictive structures themselves are the models whose workings I examine in maps, models, and understanding.

Reliably reduces entropy. Here is where the definition gets its teeth. Uncertainty has a currency: Shannon entropy, measured in bits,

\[H = -\sum_i p_i \log_2 p_i.\]

Before you learn the outcome of a fair coin flip, your uncertainty is one bit; a pattern that tells you the outcome in advance removes that bit. Knowledge is whatever removes bits — not once, by accident, but consistently across repeated and analogous situations. The reliability criterion is not a fourth condition bolted onto justified true belief. It replaces the whole apparatus: no justification clause, no truth clause, no belief clause, just a measurable track record of shrinking uncertainty.

Across branching timelines. The events knowledge predicts are events in a physical world, and on the physics I take seriously that world is the Quantum Branching Universe (QBU) — the Everettian picture in which all outcomes occur, weighted by Measure, and an agent’s uncertainty is Credence about which branch it occupies. Measure and Credence gives the primer; for this chapter, the phrase does one essential job. It fixes what “reliably” quantifies over: not a single history replayed in imagination, but the actual ensemble of branches an agent faces. A pattern constitutes knowledge when it reduces uncertainty across that ensemble — in most of the futures, by weight — not merely in the one thread of history where it happened to luck out.

Now watch the stopped clock dissolve. The clock face is a pattern, and consulting it is a procedure you can repeat — so run the reliability test. Across the ensemble of moments at which you might glance at a stopped clock, the reading reduces your uncertainty about the time by nothing: the entropy before and after the glance is the same, because the reading carries no information about the actual hour. Your 4:17 belief was true by coincidence of timing, and coincidence is precisely what the reliability criterion filters out. There is no puzzle left to patch. The Gettier industry was reverse-engineering, case by case, a distinction the entropy criterion draws in one stroke: knowledge is not belief that happens to be true, it is a pattern that reliably makes your expectations true.

Ten Stress Tests

A definition earns its keep against hard cases. Here are ten, spanning the pedestrian, the classical, and the exotic.

A weather forecast. The model says rain tomorrow, and its track record shows it is right far more often than chance. Your uncertainty about tomorrow drops by a measurable number of bits, reliably, forecast after forecast. Knowledge — the everyday kind, and the definition certifies it without ceremony.

A false belief. However sincerely held, a false belief does not reliably reduce entropy about outcomes; acting on it makes your predictions worse, not better. It fails, as it should. Note how it fails: not because a truth clause excludes it by fiat, but because falsehood shows up in the track record.

A random guess. You guess the coin will land heads, and it does. No pattern was exercised; no entropy was reduced in advance — your predictive distribution was exactly as flat after the guess as before. The lucky hit is not knowledge, and the definition never has to consult your feelings of confidence to say so.

The stopped clock. The classical Gettier case, dissolved above. Justified true belief without reliability is not knowledge. The other Gettier cases fall the same way, because they are all engineered around the same trick: a truth that arrives by a channel disconnected from the belief’s grounds. Disconnected channels have no reliability, and reliability is the criterion.

Riding a bicycle. You cannot state the control law that keeps you upright — countersteering, torque against lean angle — but your nervous system encodes it, and it reduces your uncertainty about staying vertical from guesswork to near-certainty, every ride. Tacit skill is fully paid-up knowledge on this definition, not a poor relation of the propositional kind. Any account on which knowing that is real knowledge but knowing how is mere habit has the priorities backwards; the cyclist’s cerebellum out-predicts the physicist’s blackboard on the question that matters, which is what the bicycle will do next.

A greeting ritual. Shared conventions — a language, a handshake, the rules of the queue — are patterns encoded across a community rather than a single skull, and they reliably reduce every member’s uncertainty about what the others will do. Cultural knowledge is knowledge, and the definition explains why it feels like knowledge from the inside: it makes the social world predictable.

Knowing which key opens which door. The pattern is real and reliable, and it is useless in front of any other door. All knowledge is like this; the front-door key is just the case where the conditions are too obvious to suppress. A pattern reduces entropy given a context, within a domain, under standing assumptions — knowledge inherits the conditional structure of truth itself, and a knowledge claim, like a truth claim, is compressed shorthand for a conditional.

A database no one reads. A hard drive of climate data contains correlations that would, if exercised, predict a great deal. Sitting inert, it predicts nothing. Information alone is not knowledge; it must be installed in an agent’s predictive structure, actively reducing some agent’s uncertainty. This is not a verbal stipulation but the load-bearing wall of the definition — entropy is always somebody’s uncertainty about something, and where there is no agent there is no one for the pattern to inform.

A perfectly known quantum state. Suppose you know the complete quantum state of a system about to be measured. Your epistemic uncertainty is zero — no fact about the state is hidden from you — and yet you cannot predict “the” outcome, because in the QBU every outcome occurs and the branching itself is irreducible. The definition handles this cleanly by respecting the line between Credence and Measure: knowledge eliminates the epistemic uncertainty and leaves the ontic branching untouched, exactly as it should. A definition that demanded prediction of single outcomes here would be demanding clairvoyance about a fact that does not exist.

Steering across branches. Conversely, an agent whose patterns tell it which actions lead to favorable branch-weights — which choices put most of its future Measure in branches worth occupying — is exercising knowledge in the fullest sense. This is what knowledge is ultimately for in a branching world: not passive anticipation but navigation.

Ten scenarios, no epicycles. The definition certifies the forecast, the bicycle, the handshake, the key, and the branch-navigator; it rejects the false belief, the lucky guess, the stopped clock, and the unread database; and it draws the epistemic-ontic line exactly where the physics draws it.

Four Kinds of Knowledge

One definition, then — but not one kind. Patterns can reduce entropy in structurally different ways, and running the kinds together produces some of epistemology’s most persistent confusions. Four categories cover the territory.

Explanatory knowledge consists of general theories — quantum mechanics in its many-worlds form, evolution by natural selection, general relativity. These are frameworks that hold across every branch within their explanatory domain: there is no timeline where evolution is false and creationism true, because the theory’s content is not a fact about which branch you are in at all. Accordingly, explanatory theories are evaluated by explanatory coherence, simplicity, generality, and resistance to criticism — the critical-rationalist virtues — and the interesting question of whether they can also bear credences is one I settle in defense of Bayes.

Empirical knowledge is knowledge of which branch you are in: whether you carry the genetic variant, whether it will rain tomorrow, what a historical actor’s motives were. This is the native domain of Credence and Bayesian updating — probabilistic through and through, with each new observation trimming the set of timelines consistent with your evidence. The taxonomy of what, exactly, an agent can be uncertain about is richer than it first appears, and I map it in the varieties of uncertainty.

The explanatory-empirical line is worth drawing sharply because each side’s method fails on the other’s territory. Judging a universal explanatory framework as if it were a weather forecast — the misuse of Bayes that Deutsch and his followers rightly attack — is one error; treating a branch-specific empirical question as if criticism alone could settle it, with no probabilistic bookkeeping, is the mirror error. Most of the shouting between Bayesians and critical rationalists is a failure to notice that they are custodians of different categories.

Sharply drawn is not exhaustively drawn, though, and the boundary has honest hybrid cases. A parameterized theory — general relativity plus a Hubble constant, particle physics plus a dark-matter density — is an explanatory framework wrapped around an empirical dial: the framework answers to criticism while the parameter answers to Bayes. Historical interpretation blends the two the other way, running general explanatory machinery (economics, sociology, evolutionary psychology) over irreducibly branch-specific gaps in the record. The categories are joints in the territory, not a demand that every specimen fall on one side.

Formal knowledge — mathematics and logic — is non-empirical, necessary, and derivable a priori: arithmetic, set theory, Gödel’s incompleteness theorems. No branch of the multiverse makes seven prime and no observation was ever needed to establish it. Its entropy reduction is real but distinctive: it removes uncertainty about the consequences of premises, which is why it multiplies the power of the other three kinds — the forecast model is empirical patterns run through formal machinery.

Tacit knowledge — the bicycle, the violin, the seasoned diagnostician’s hunch — is embodied, implicit, and resistant to explicit formulation. It passed the stress test above on the same terms as everything else, and that is the point of listing it as a full category rather than a footnote: the definition never required knowledge to be statable, only to be encoded and reliable.

The typology also explains why the classical definition was doomed. Justified true belief was reverse-engineered from one category — explicit, propositional, single-timeline empirical claims — and then declared the essence of the whole. Tacit knowledge fails its belief clause, formal knowledge trivializes its truth clause, explanatory knowledge strains its justification clause, and Gettier cases break it even on its home turf. Entropy reduction covers all four kinds without strain because it is a claim about what knowledge does, not about the format it comes in.

Knowledge Cashed Out

This account slots into the volume’s larger architecture. Beliefs, as I argue in what beliefs are, are predictive models, and their virtue is calibration; knowledge is the portion of an agent’s predictive structure whose calibration is earned — patterns with a track record of paying out in bits. Truth is conditional validity relative to specified assumptions; knowledge is conditional too, a pattern that delivers within its domain and is silent outside it, front-door key writ large.

And the definition returns an answer to the oldest question about knowledge — what is it worth? — in the only currency an agent ultimately spends. Entropy about the future is risk: the predator not anticipated, the market not modeled, the side effect not foreseen. A pattern that reliably removes bits of that uncertainty is not a representation to be admired; it is traction. The measure of what you know is what you can predict, and the measure of what you can predict is how deliberately you can steer — through the one world of common experience, and through the branching futures beneath it.